Protein-Protein Binding Affinity Prediction Based on Wavelet Package Transform and Two-Layer Support Vector Machines

Precisely inferring the affinities of protein-protein interaction is essential for evaluating different methods of protein-protein docking and their outputs and also opens a door to inferring real status of cellular protein-protein complex. Accumulation of measured affinities of determined protein complex structures with high resolution facilitate the realization of this ambitious goal. Previous physical model based scoring functions failed to predict the affinities of diverse protein complexes. Therefore, accurate method for binding affinity prediction is still extremely challenging. Machine learning methods are promising to address this problem. However, current machine learning methods are not compatible to this task, which obstructs the effective application of these methods. We propose a Wavelet Package Transform (WPT) combined with two-layer support vector regression (TLSVR-WPT) model to implicitly capture binding contributions that are hard to model explicitly. Wavelet package transform greatly reduced the dimension of input features into machine learning model. The TLSVR circumvents both the descriptor compatibility problem and the need for problematic modeling assumptions. Input features for TLSVR first layer are eight features transformed by Wavelet Transform Package from scores of 2209 interacting atom pairs within each distance bin. The output of the first layer is combined by the next layer to infer the final affinities. A satisfactory result of R = 0.81 and SD = 1.40 was achieved when 2209 features were reduced to eight ones by 3-level Wavelet Package Transform. Results demonstrate that wavelet package transform greatly reduced the dimension of the input features into SVR without reducing the accuracy in predicting the protein binding affinity.